Main Article Content
Medication adherence is associated with a reduction of adverse outcomes in heart failure (HF). However, this association is complex to estimate accurately because adherence (exposure) can vary during the follow-up period. Adherence can be estimated as a fixed exposure to predict outcomes, and this is known as a landmark analysis. In contrast, adherence can also be estimated as a dynamic exposure which varies over time in the follow-up period. This is known as a time-varying analysis and is expected to be the more precise method.
Objectives and Approach
We compared these two methods in a HF cohort. We identified a population-based cohort of 3619 heart failure patients, aged 65-84 years hospitalised in Western Australia from 2003-2007 and who survived to 1-year post-discharge (landmark date). Adherence to renin-angiotensin system inhibitors (RASI) and β-blockers was calculated using proportion of days covered (PDC) expressed either as a fixed time exposure (in landmark analysis) or a varying exposure (in time-dependent analysis). The latter was updated every 30 days after the landmark date. Cox regression models were used to investigate the association between adherence and all-cause death at 1- and 3-years post-landmark date.
For 1-year outcomes, hazard ratios (HR) for every 10% increase in PDC were similar between models from landmark analyses (RASI adherence: 0.93, 0.90-0.97; β-blocker adherence: 0.96, 0.92-1.0) and time-dependent analyses (RASI adherence: 0.94, 0.91-0.97; β-blockers adherence: 0.95, 0.92 -0.99). However, 95% confidence intervals estimated from time-dependent models were narrower than those from landmark analyses. HRs were slightly closer to the null when estimated from time-dependent models. A similar pattern was seen with 3-year outcomes.
Conclusion / Implications
Time-dependent analysis of adherence-outcome associations results in more precise estimates of hazard ratios. Estimates of HRs from landmark analysis models were similar but usually lower than those from time-dependent models.
This work is licensed under a Creative Commons Attribution 4.0 International License.